This study aims to contribute to the scarce data available about the abilities of untrained lay persons to perform hands-only cardio-pulmonary resuscitation (CPR) on a manikin and the improvement of their skills during training with an autonomous CPR feedback device. The study focuses on the following questions: (i) Is there a need for such a CPR training device? (ii) How adequate are the embedded visual feedback and audio guidance for training of lay persons who learn and correct themselves in real time without instructor guidance? (iii) What is the achieved effect of only 3 min of training? This is a prospective study in which 63 lay persons (volunteers) received a debriefing to basic life support and then performed two consecutive 3 min trials of hands-only CPR on a manikin. The pre-training skills of the lay persons were tested in trial 1. The training process with audio guidance and visual feedback from a cardio compression control device (CC-Device) was recorded in trial 2. After initial debriefing for correct chest compressions (CC) with rate 85-115 min(-1), depth 3.8-5.4 cm and complete recoil, in trial 1 the lay persons were able to perform CC without feedback at mean rate 95.9 ± 18.9 min(-1), mean depth 4.13 ± 1.5 cm, with low proportions of 'correct depth', 'correct rate' and 'correct recoil' at 33%, 43%, 87%, resulting in the scarce proportion of 14% for compressions, which simultaneously fulfill the three quality criteria ('correct all'). In trial 2, the training process by the CC-Device was established by the significant improvement of the CC skills until the 60th second of training, when 'correct depth', 'correct rate' and 'correct recoil' attained the plateau of the highest quality at 82%, 90%, 96%, respectively, resulting in 73% 'correct all' compressions within 3 min of training. The training was associated with reduced variance of the mean rate 102.4 ± 4.7 min(-1) and mean depth 4.3 ± 0.4 cm, indicating a steady CC performance achieved among all trained participants. Multivariable linear regression showed that the compression depth, rate and complete chest recoil did not strongly depend on lay person age, gender, height, weight in pre-training and training stage (correlation coefficient below 0.54). The study confirmed the need for developing CPR abilities in untrained lay persons via training by real-time feedback from the instructor or CC-Device. The CC-Device embedded feedback was shown to be comprehensible and easy to be followed and interpreted. The high quality of the CC-Device-assisted training process of lay persons was confirmed. Thus learning or refresher courses in basic life support could be organized for more people trained at the same time with fewer instructors needed only for the initial debriefing and presentation of the CC-Device.

Deep neural networks (DNN) are state-of-the-art machine learning algorithms that can be learned to self-extract significant features of the electrocardiogram (ECG) and can generally provide high-output diagnostic accuracy if subjected to robust training and optimization on large datasets at high computational cost. So far, limited research and optimization of DNNs in shock advisory systems is found on large ECG arrhythmia databases from out-of-hospital cardiac arrests (OHCA). The objective of this study is to optimize the hyperparameters (HPs) of deep convolutional neural networks (CNN) for detection of shockable (Sh) and nonshockable (NSh) rhythms, and to validate the best HP settings for short and long analysis durations (2–10 s). Large numbers of (Sh + NSh) ECG samples were used for training (720 + 3170) and validation (739 + 5921) from Holters and defibrillators in OHCA. An end-to-end deep CNN architecture was implemented with one-lead raw ECG input layer (5 s, 125 Hz, 2.5 uV/LSB), configurable number of 5 to 23 hidden layers and output layer with diagnostic probability p ∈ [0: Sh,1: NSh]. The hidden layers contain N convolutional blocks × 3 layers (Conv1D (filters = Fi, kernel size = Ki), max-pooling (pool size = 2), dropout (rate = 0.3)), one global max-pooling and one dense layer. Random search optimization of HPs = {N, Fi, Ki}, i = 1, … N in a large grid of N = [1, 2, … 7], Fi = [5;50], Ki = [5;100] was performed. During training, the model with maximal balanced accuracy BAC = (Sensitivity + Specificity)/2 over 400 epochs was stored. The optimization principle is based on finding the common HPs space of a few top-ranked models and prediction of a robust HP setting by their median value. The optimal models for 1–7 CNN layers were trained with different learning rates LR = [10−5; 10−2] and the best model was finally validated on 2–10 s analysis durations. A number of 4216 random search models were trained. The optimal models with more than three convolutional layers did not exhibit substantial differences in performance BAC = (99.31–99.5%). Among them, the best model was found with {N = 5, Fi = {20, 15, 15, 10, 5}, Ki = {10, 10, 10, 10, 10}, 7521 trainable parameters} with maximal validation performance for 5-s analysis (BAC = 99.5%, Se = 99.6%, Sp = 99.4%) and tolerable drop in performance (<2% points) for very short 2-s analysis (BAC = 98.2%, Se = 97.6%, Sp = 98.7%). DNN application in future-generation shock advisory systems can improve the detection performance of Sh and NSh rhythms and can considerably shorten the analysis duration complying with resuscitation guidelines for minimal hands-off pauses.

Minimum ''hands-off'' intervals during cardiopulmonary resuscitation (CPR) are required to improve the success rate of defibrillation. In support of such life-saving practice, a shock advisory system (SAS) for automatic analysis of the electrocardiogram (ECG) contaminated by chest compression (CC) artefacts is presented. Ease of use for the automated external defibrillators (AEDs) is aimed and therefore only processing of ECG from usual defibrillation pads is required. The proposed SAS relies on assessment of outstanding components of ECG rhythms and CC artefacts in the time and frequency domain. For this purpose, three criteria are introduced to derive quantitative measures of band-pass filtered CC-contaminated ECGs, combined with three more criteria for frequency-band evaluation of reconstructed ECGs (rECG). The rECGs are derived by specific techniques for CC waves similarity assessment and are reproducing to some extent the underlying ECG rhythms. The rhythm classifier embedded in SAS takes a probabilistic decision designed by statistics on the training dataset. Both training and testing are fully performed on real CC-contaminated strips of 10 s extracted from human ECGs of out-ofhospital cardiac arrest interventions. The testing is done on 172 shockable strips (ventricular fibrillations VF), 371 nonshockable strips (NR) and 330 asystoles (ASYS). The achieved sensitivity of 90.1% meets the AHA performance goal for noise-free VF (>90%). The specificity of 88.5% for NR and 83.3% for ASYS are comparable or even better than accuracy reported in literature. It is important to note that, the aim of this SAS is not to recommend shock delivery but to advice the rescuers to ''Continue CPR'' or to ''Stop CPR and Prepare for Shock'' thus minimizing ''hands-off'' intervals.

No abstract

This paper presents a bench study on a commercial automated external defibrillator (AED). The objective was to evaluate the performance of the defibrillation advisory system and its robustness against electromagnetic interferences (EMI) with central frequencies of 16.7, 50 and 60 Hz. The shock advisory system uses two 50 and 60 Hz band-pass filters, an adaptive filter to identify and suppress 16.7 Hz interference, and a software technique for arrhythmia analysis based on morphology and frequency ECG parameters. The testing process includes noise-free ECG strips from the internationally recognized MIT-VFDB ECG database that were superimposed with simulated EMI artifacts and supplied to the shock advisory system embedded in a real AED. Measurements under special consideration of the allowed variation of EMI frequency (15.7-17.4, 47-52, 58-62 Hz) and amplitude (1 and 8 mV) were performed to optimize external validity. The accuracy was reported using the American Heart Association (AHA) recommendations for arrhythmia analysis performance. In the case of artifact-free signals, the AHA performance goals were exceeded for both sensitivity and specificity: 99% for ventricular fibrillation (VF), 98% for rapid ventricular tachycardia (VT), 90% for slow VT, 100% for normal sinus rhythm, 100% for asystole and 99% for other non-shockable rhythms. In the presence of EMI, the specificity for some non-shockable rhythms (NSR, N) may be affected in some specific cases of a low signal-to-noise ratio and extreme frequencies, leading to a drop in the specificity with no more than 7% point. The specificity for asystole and the sensitivity for VF and rapid VT in the presence of any kind of 16.7, 50 or 60 Hz EMI simulated artifact were shown to reach the equivalence of sensitivity required for non-noisy signals. In conclusion, we proved that the shock advisory system working in a real AED operates accurately according to the AHA recommendations without artifacts and in the presence of EMI. The results may be affected for specificity in the case of a low signal-to-noise ratio or in some extreme frequency setting.